WeatherNext 3 Reality Check: 5 km Is Not Every Variable, Brightband Skips Rain, and 60% Needs a Target

Google DeepMind and Google Research released WeatherNext 3 on September 3, 2026 as their new operational global AI weather model. It is a real technical step forward: the system ingests low-latency geostationary satellite mosaics, can initialize a forecast every hour, produces a 64-member probabilistic ensemble, and adds much finer surface outputs than WeatherNext 2.

But several launch headlines compress different resolutions, datasets and evaluation targets into single numbers. The safest way to read WeatherNext 3 is to keep five things separate: what is actually 5 km, which runs extend to 15 days, what the 50–60% precipitation claims compare against, what Brightband independently evaluates, and which Google Cloud surfaces expose WeatherNext 3 versus WeatherNext 2.

The “5 km” claim applies to selected station-calibrated variables

WeatherNext 3 is not a single 5 km global atmospheric grid.

Google’s developer specification lists three main output scales:

• about 0.05° (~5 km) for the station-calibrated 2 m temperature and dew-point heads;
• about 0.1° (~10–11 km) for core gridded surface variables such as wind, pressure, precipitation, cloud cover and solar radiation;
• about 0.25° (~25 km) for many pressure-level atmospheric variables.

The research paper explains why. WeatherNext 3 uses separate encoders and decoders for different native resolutions, maps them into a shared Functional Generative Network mesh transformer, and then decodes each modality at its appropriate resolution. The station head can be queried continuously in space and time using metadata such as elevation and land/sea mask.

That makes the five-kilometre figure genuine, but it should not be generalized to every forecast field. A developer evaluating wind aloft, precipitation or a pressure-level variable should use the documented resolution for that exact field instead of assuming the headline grid applies universally.

Hourly initialization does not mean every hourly run is a 15-day forecast

WeatherNext 3 can initialize 24 times per day, but the forecast horizon depends on the initialization.

Google’s current model guide says the four main 00, 06, 12 and 18 UTC cycles extend to 360 hours (15 days). The interim hourly cycles between those main runs currently extend to 48 hours.

This distinction matters for products that advertise “hourly forecasts out to 15 days.” The model outputs hourly timesteps, and it can start a new run each hour, but not every hourly initialization has the full 15-day horizon in the published operational contract.

The paper also says the system still uses a six-hour outer autoregressive step internally. Some variables are predicted at finer hourly offsets inside those windows. That architectural detail helps explain one of the model’s documented failure modes: certain non-autoregressive output heads can show jumps at six-hour boundaries even though the published data are hourly.

The precipitation numbers are real, but the target changes the result

Google’s launch article reports three different precipitation improvements for medium-range forecasts at early lead times:

• up to 60% CRPS improvement against IMERG, NASA’s satellite precipitation product;
• about 30% against MRMS, the U.S. Multi-Radar/Multi-Sensor precipitation analysis;
• about 10% against rain gauges.

Those are not three measurements of the same ground truth. The WeatherNext 3 paper is explicit that precipitation evaluation is difficult because no single precipitation dataset is perfect.

IMERG is global and convenient, but it is also used as a WeatherNext 3 training target, so it is not an independent ground truth for the IMERG output head and could indirectly influence the jointly trained PARDIG head. MRMS is higher-quality radar-based evidence but covers primarily the contiguous United States. Rain gauges are direct observations but are geographically sparse and unevenly distributed.

Google’s developer guide separately summarizes precipitation performance as up to a 50% reduction in Brier score and CRPS compared with numerical-weather-prediction baselines. That should not be silently substituted for the 60% IMERG number: the wording, metric set, baselines and comparison slices are different.

The practical lesson is simple: “60% better rain forecast” is too broad. The strongest launch percentage belongs to a specific probabilistic metric and a specific satellite-derived verification target.

Brightband is useful independent evidence — but it currently skips WeatherNext 3 precipitation

Google cites Brightband’s Operational WeatherBench as independent live evidence that WeatherNext 3 is leading operational global models.

Brightband’s methodology is substantially more informative than a one-number leaderboard. It scores models using WeatherBench-X, separates deterministic ensemble-mean evaluation from probabilistic ensemble-member evaluation, and records model, target, variable, metric, region, initialization and lead time separately.

For WeatherNext 3, Brightband currently verifies its surface fields against ECMWF IFS analysis at 0.1°. It also explicitly states that it does not currently publish WeatherNext 3 precipitation scores. WeatherNext 3 precipitation is trained toward satellite-based targets, and Brightband says downscaling or regridding it into the common precipitation comparison would create a difficult apples-to-oranges evaluation.

That means Brightband can independently support a claim that WeatherNext 3 is highly competitive on the variables it actually scores, but it cannot independently validate Google’s 60% precipitation headline. Those are two different evidence tracks and should stay separate.

Brightband also samples 16 members when computing its probabilistic metrics, using finite-ensemble corrections intended to make the estimates comparable across ensemble sizes. That methodology is more rigorous than treating one ensemble mean as the whole system.

Google’s own paper uses both a full-year research evaluation and a short operational window

The main research evaluation uses the full year of 2024 with a model trained through the end of 2023. That helps reduce leakage from the test year.

The paper also includes a quasi-real-time operational comparison for July 1 to August 11, 2026, using the production WeatherNext 3 checkpoint trained through June 30, 2026. Google calls the six-week sample relatively small and warns against over-interpreting smaller differences.

In that six-week comparison, the paper reports WeatherNext 3 outperforming ECMWF AIFS ENS v2 across all evaluated upper-level variables, with roughly 10% average improvement during the first forecast week, while also showing strong gains on station temperature, humidity and wind-speed evaluations. The authors note, however, that some single-level comparisons are affected by resolution, interpolation and verification-target choices.

That caveat is important. Weather model ranking can change with variable, region, lead time, verification dataset and interpolation method. A universal “best weather model” label hides those dimensions.

The model has visible spatial and temporal artifacts

WeatherNext 3 is trained primarily to optimize marginal probabilistic skill, and the paper says that can produce unrealistic covariance structure in individual ensemble members.

The researchers document hexagonal mesh artifacts, especially in precipitation and station outputs. They also show temporal discontinuities across six-hour boundaries in the station head and some precipitation heads, plus a per-member warm-or-cold global bias that can reset when a new noise vector is drawn.

The ensemble median and other marginal statistics reduce many of these artifacts, which is why the model can still be useful for probabilistic applications. But this is a reminder that a low CRPS does not guarantee every individual forecast trajectory is physically smooth or visually realistic.

Google’s developer documentation repeats these warnings and says WeatherNext 3 must not be used as the sole authority for life-safety decisions or official warnings.

“Operational” does not mean open weights or unrestricted WeatherNext 3 inference

WeatherNext 3 forecast data are available through Google Cloud Storage, BigQuery and Earth Engine, with access controls and experimental-data terms. The Earth Engine catalog explicitly warns that the operational model’s accuracy may not exactly match the research paper’s reported accuracy.

There is another product boundary worth preserving: Google’s current managed-inference specification still identifies WeatherNext 2 as the model exposed for customizable managed generation. WeatherNext 3 is available as operational forecast data, but the public documentation checked in this review does not establish a generally available custom WeatherNext 3 inference endpoint with public per-run pricing.

So “WeatherNext 3 is on Google Cloud” should not be interpreted as “developers can invoke the exact WeatherNext 3 production checkpoint on demand with arbitrary ensemble sizes.”

Public price-per-forecast, parameter count and self-hostable production weights were not established in the primary documentation checked here. Those facts remain unknown rather than being inferred from older WeatherNext releases.

Independent feedback is positive, but the “first raw-observation model” claim is contested

TechCrunch quoted Brightband atmospheric scientist Daniel Rothenberg positively on the value of directly predicting station measurements, because it connects model output more closely to what is observed at specific locations.

The same reporting records a competitive disagreement over Google’s “first” framing. WindBorne says its WeatherMesh 6 had already incorporated raw observations from weather balloons and other sources since late 2025. That does not negate WeatherNext 3’s capabilities, but it makes “first AI weather model to use raw observations” too broad without defining resolution, coverage, input types and operational status.

Public Reddit reaction after the launch was generally enthusiastic about practical forecasting improvements, but the discussion also included users reporting poor prior Google-weather accuracy in their own areas and asking whether the new model was actually active in their region. Those are self-selected anecdotes, not measured reliability evidence.

I did not find sufficiently attributable, reproducible X posts from independent meteorological researchers during this bounded pass to support an X-consensus claim, so none is invented.

SWE-bench Verified and SWE-bench Pro do not apply here

SWE-bench Verified and SWE-bench Pro are software-engineering benchmarks. WeatherNext 3 is a probabilistic weather-forecasting system, not a coding agent, and no WeatherNext 3 score on either benchmark is relevant to this release.

The appropriate evaluation families are weather-specific: CRPS, RMSE, Brier score, anomaly correlation, spread-skill diagnostics, cyclone errors, station-observation comparisons and operational live backtesting.

The same benchmark-hygiene rule still applies: preserve the exact model version, initialization, ensemble size, target dataset, interpolation method, time window, region, metric and lead time before ranking systems.

Practical takeaway

WeatherNext 3 is a substantial operational AI-weather release because it combines more recent satellite observations, hourly initialization, finer surface outputs, 64-member probabilistic forecasts and direct station/precipitation targets in one system.

The most defensible interpretation is narrower than the marketing shorthand:

• 5 km applies to selected station-calibrated surface variables, not every atmospheric field.
• 15 days applies to the main six-hourly cycles; interim hourly starts currently run to 48 hours.
• 60% precipitation improvement is a scoped CRPS result against IMERG, not a universal rain-accuracy number.
• Brightband’s independent leaderboard does not currently score WeatherNext 3 precipitation, so it cannot be used as external confirmation of that 60% claim.
• The paper’s operational comparison spans only six weeks and explicitly warns against over-reading small differences.
• Individual ensemble members can show mesh and six-hour-boundary artifacts.
• WeatherNext 3 forecast data are operationally accessible, but current managed custom inference documentation still names WeatherNext 2.
• Official national meteorological warnings remain the authority for life-safety decisions.

For AI evaluation more broadly, WeatherNext 3 is a useful case study in why a model release should be reported as a matrix of model version, target, horizon, resolution, metric and access mode—not as one headline percentage.